This action might not be possible to undo. Are you sure you want to continue?
Nonlinear Diffusion and Martingale Option Pricing
J.L. McCauley+, G.H. Gunaratne++, and K.E. Bassler Physics Department University of Houston Houston, Tx. 77204 firstname.lastname@example.org Senior Fellow COBERA Department of Economics J.E.Cairnes Graduate School of Business and Public Policy NUI Galway, Ireland
Institute of Fundamental Studies Kandy, Sri Lanka
Key Words: Markov process, scaling, Hurst exponent, option pricing, BlackScholes, Martingales, fat tails Abstract First, classes of Markov processes that scale exactly with a Hurst exponent H are derived in closed form. A special case of one class is the Tsallis density, advertised elsewhere as ‘nonlinear diffusion’ or ‘diffusion with nonlinear feedback’. But the Tsallis model is only one of a very large class of linear diffusion with a student-t like density. Second, we show by stochastic calculus that our generalization of the Black-Scholes partial differential equation (pde) for variable diffusion coefficients is equivalent to a Martingale in the risk neutral discounted stock price. Previously, this was proven for the restricted case of Gaussian logarithmic returns by Harrison and Kreps, but we prove it here for large classes of empirically useful and theoretically interesting returns models where the diffusion coefficient D(x,t) depends on both logarithmic returns x and time t. Finally, we prove that option prices blow up if fat tails in returns x are included in the market distribution.
We begin with the fact that a Hurst exponent H describes. t)dx = ct 2H $ ! To obtain variance scaling (1) we need a probability density f(x. (1) where c is a constant. t)dB(t) (3) ! . Initially. without autocorrelations in x(t).2 Markov processes with scaling A Markov process [3. 1. t) + D(x. Here. the initial point in the time series is assumed to be xo=0 at t=0. t) = t "H F(u). and also with fractional Brownian motion. Markov processes with scaling densities 1.t) that scales as  f(x.1. u = x/t H . the Hurst exponent taken alone cannot be used to determine the existence of long term memory. Any stochastic differential equation (‘Langevin equation’) dx = R(x.1 Hurst exponents Hurst exponents are normally asumed to imply auto-correlations that describe long term memory.4] is a stochastic process without memory. without other assumptions . the scaling of the variance of a stochastic process x(t). #$ " 2 = % x2 f(x. We will explain why H≠1/2 is consistent both with Markov processes. (2) ! Up to this point we have not assumed any underlying stochastic dynamics. Our discussion so far is consistent with both fractional Brownian motion (fBm)  and a Markov process . where the autocorrelations extend over arbitrarily long nonoverlapping time intervals. Therefore. we limit our analysis to a drift-free process so that <x>=0. which by construction have no autocorrelations.
t) depends on both x and t.t) also do not scale. and arbitrary choices of R(x. From the sde (3) we can also calculate the variance as " 2 = # ds # dxf(x.t)/2 do not scale. B(t) is the usual Wiener process.t) because that only these solutions exhibit exact scaling (1. Drifts of the form R(x.0)=f(x. then we obtain  2H(uF(u))" + (D(u)F(u))"" = 0 (6) ! .t) (as opposed to functionals) as drift and diffusion coefficients generates a Markov process [2. and Ito calculus  is assumed. u = x/t H (4) follows from combining (1) and (2) with (1b). (5) ! which can be rewritten to include a constant drift rate R by replacing x by xRt .6]. The full Green function g(x. s).t. We restrict ourselves here to the case where g(x.7] "f 1 "2 = (Df) "t 2 "x2 .xo.4).2.with functions R(x.to) is needed in principle for risk neutral option pricing  but doesn’t scale  and also can’t be calculated in closed analytic form when D(x. ! The sde (3) with R=0 generates the Green function g of the Fokker-Planck pde [2.t) and D(x. and also because the density f(x.9]. Inserting the scaling forms (2) and (3) into the pde (5).t) is what one obtains directly from histograms of finance market returns data [7. t) = t 2H"1 D(u). Scale invariance is generally broken by a drift rate depending on x.5. s)D(x. 0 $% t % (1b) so that diffusion coefficient scaling ! D(x.t. For the time being we take R=0.0.t)=µ-D(x.
0<H<1. does not imply aucorrelations Consider the Markov process (3) in its integrated form ! "x(t. 3. with µ=2+2H/ε. So we now know where student-tlike distributions come from dynamically. then H≠1/2 does not imply .whose solution is given by F(u) = C "2H # udu/D(u) . (8) This choice generates the two parameter class of student-t.t+T)=x(t+T)x(t) over the two nonoverrlapping time intervals . Δx(t. and also obtain a finite variance scaling as σ2=ct2H whenever µ>3. where H and ε are independent parameters to be determined empirically.t-Τ)=x(t)-x(t-T). In that case. t + T) = # t t +T D(x(s).4) holds with H≠1/2 . we study the class of quadratic diffusion coefficients D(u) = d"(#)(1 + #u 2 ) . With d’(ε)=1 we obtain the generalization of our earlier prediction  to arbitrary H. We can generate both the exponential middle  and the fat tailed extremes systematically by using a diffusive model. meaning that x(t+T)x(t)≠x(T) depends both on Τ and t.3 H≠1/2. we can generate all fat tail exponents 2<µ<∞ (2<µ<7 is observed ). taken alone. For large u this model fits the fat tails in financial data for all times t . If the nonstationary stochastic process x(t) has nonstationary increments. s)dB(s) (10) ! Even when scaling (1. e D(u) (7) This generalizes our earlier results  to include H≠1/2.like densities ! F(u) = C"(1 + #u 2 )$1$H/#d"(#) (9) with fat tail exponent µ=2+2H/εd’(ε). then there can be no autocorrelation in the increments Δx(t. ! Next.2.
So. generates a linear diffusion pde.s) dB(s) 0 t +T t H"1/2 t = # t T 0 t +T D(x(s). but apparently none of the authors noted the following: a Langevin eqn. In the literature these two separate requirements are often confused together into a soup. but a nonlinear diffusion eqn. As always with stochastic equations. When the increments are nonstationary.5]. with ordinary functions as drift and diffusion coefficients is Markovian. The argument that H≠1/2 implies long time correlations fails for Markov processes (and for the finance market as well.4.s) dB(s) " # D(x(s). then performing a correct data analysis nontrivial . x(t+T)-x(t)=x(T). The Green function. and (ii) the variance must scale. <x2(t)>=ct2H . (11) = # s+t D(x/ s + t ) dB(s) $ x(T) ! Direct calculation of the autocorrelation function  shows that it vanishes independently of the value of H. Given the Tsallis density  . has no Green function.4) describes a nonstationary process with nonstationary increments whenever H≠1/2: x(t+ T) " x(t) = # 0 t +T D(x(s). equality is understood as ‘equality in distribution’. We can clear up the confusion.7]. either the Tsallis model is nonlinear but has no underlying Langevin description.8 The Tsallis Model is not generated by ‘nonlinear diffusion’ Many papers have been written claiming that the Tsallis density is generated by both a ‘nonlinear Fokker-Planck pde’ and by a Langevin eqn. the usual Fokker-Planck pde [2. [13-17] as well.s) dB(s) = # s H"1/2 H D(u) dB(s) .autocorrelations . or else the Tsallsi density is a Markov process in a nonlinear disguise. to zeroth order) precisely because the stochastic integral (10) with the scaling forms (1. or condidional probability density.5. To obtain the long time autocorrelations of fractional Brownian motion (fBm)  two independent conditions must be satisfied: (i) the time series must have stationary increments. 3.2. is required to define a Markov process [4.
F(u)=CD(u)-1-H/εd’(ε). A similar nonlinear disguise is possible for our entire two-parameter studentt-like class solutions (9). but rewriting (19) as a nonlinear pde in the case of quadratic diffusion superficially masks the real Markovian nature of the dynamics.fq (x.t) given by (8) reduces exactly to (14) if H=1/(3-q). f(x.8). Because the Tsallis density (11) describes Markovian dynamics. it cannot . The Tsallis model definition  1"q D q (x.16) with (9.’ [13-17] is really a linear pde disguised superficially as a nonlinear one. because for quadratic diffusion D(u)=d’(ε)(1+εu2). t) = (c(2 " q)(3 " q))"H t "H (1 + (q " 1)x2 /C 2 (q)t 2H )1/(1"q ) (11) with H=1/(3-q). is µ=2/(q1). t) (14) yields ! D q (x. and H=1/(3-q). where ! C(q) = c(q"1)/2(3"q) ((2 " q)(3 " q))H (12) ! with c a constant then the fat tail exponent. the solution of the Fokker-Planck pde (5) is a power of the diffusion coefficient. t) = (c(2 " q)(3 " q))2H"1 t 2H"1 (1 + (q " 1)x2 /C 2 (q)t 2H ) which we can rewrite as (15) ! D q (x. All of these solutions also trivially satisfy a nonlinear pde. we need only write ε=(q-1)/C2(q) and d’(ε)=d(q).t) ≈ x-m for x>>1. t) = fq (x. This means that the Tsallis density fq actually derives from the linear diffusion pde 2 "f 1 " (D q f ) = (17) "t 2 "x2 ! so that the so-called ‘nonlinear Fokker-Planck eqn. t) = d(q)t 2H"1 (1 + ((q " 1)/C 2 (q))u 2 ) = t 2H"1 D q (u) (16) ! To compare (11. Our Fokker-Planck-generated density f(x.
H=1/(3-q)≠1/2 merely signals that the increments x(t) are nonstationary. if we choose µ=r in the latter .describe long-time correlated signals like fBm. t) "2 u(x. t. t)/ 2) )+ "t "x 2 "x2 (21) ! for the market Green function of returns g.t)/2 included [7. corresponding to the FokkerPlanck pde (the second Kolmogorov equation) "g(x. t o ) = #((µ # D(x. t) ru(x. t) D(x. we explain why nonlinear diffusion would be inconsistent with risk neutral option pricing: the Green function for the Black-Scholes pde solves our market Fokker-Planck pde (5) with a nonscaling drift term R(x. xo . t) "2 v 0= + (r # D(x. xo . t) = + (r # D(x. t)/ 2) + "t "x 2 "x2 . xo . t o ) "g(x.9]. This pde is exactly the backward time equation. Here. (20) ! This is a very beautiful result. With the choice µ=r then both pdes are solved by the same Green . t) "u(x. t) "2 g(x. Pricing options via Martingales 2.t)=µD(x. t o ) D(x. or first Kolmogorov equation . t.9] ! "u(x. 2. t)/ 2) + "t "x 2 "x2 becomes (19) ! "v "v D(x. t. Next.2 The Black-Scholes PDE and Kolmogorov’s First PDE With the transformation u = er (t "T) v (18) the generalized Black-Scholes pde (see  for the original Gaussian returns model) in the returns variable [7.
2.22]. t)dxT "# # . e. t o ) D(xo .t.function g. Why fat tails are ignored is explained below. and K is the strike price) and pc is the consensus price.t) in the drift: we’ve priced options in agreement with traders prices by treating R as constant while ignoring fat tails. T " t) = er (t "T) $ (p T " K)%(p T " K)g(xT .to)=g(x.to) solves the Fokker-Planck pde (21) in (x. (23) ! where xT=lnpT/pc and x=lnp/pc where p is the price at present time t (T is the expiration time.t) to the drift R(x. T. The forward time Kolmogorov pde . This is like physics: everything can be calculated when one knows the Green function [7. t o ) 0= + (r # D(xo . t o ) "v(xo . We can.t. K. is unfortunately insensitive to whether we keep or ignore the term D(x. the forward and backward time pdes are adjoints of each other .21]. using only the exponential density of returns [7..g. t o )/ 2) + 2 "t o "xo 2 "xo (22) where v(xo.xT. use the market Green function g to price calls risk neutrally as ! C(p.T) of the pde (20). The Fokker-Planck operator cannot be made self-adjoint via boundary conditions. option pricing is not a strong test of the underlying market dynamics. It’s easy to check that the diffusive contribution D(x. x. where the dagger denotes the adjoint. However.t). according to the theory of backward time integration [19. or ‘value’ [10.xo.20] we must understand (20) as "v(xo .t)=rD(x.9]. so that no information is provided by solving the option pricing pde (19) that is not already contained in the Green function of the Market Fokker-Planck equation (21). To be explicit. t o ) "2 v(xo .3 Generalized Black-Scholes describes a Martingale The call price is calculated from the Green function v=g+(x.t)/2 breaks Hurst exponent scaling.
t)=E(y."g " "2 D(xT . That the Black-Scholes pde is equivalent to a Martingale in the risk neutral discounted stock price was proven abstractly for the case of the Gaussian returns model .t).t)=d(p. to requiring that the variance is finite.t)=D(x.t. Then with g(x.x. t)dB (25) ! where d(p.yo.to)=G(y. Our proof is not restricted to the unphysical assumption that D(x. t) = # (# G) + 2 ( G) "t "y 2 "y 2 with E(y.t)=D(x. T) =# ((r # D(xT .t) is independent of x. The price sde corresponding to this Fokker-Planck pde (dropping subscripts capital T.T).t) and r is the risk neutral rate of return. t)dB(t) . which corresponds to the sde (28) ! dy = "E(y.to) (since dx=dy)  we obtain E(y. t) "G " "2 E(y. for convenience) is ! dp = rpdt + p 2 d(p.t)=D(x. in adition.t.t).t. t)dt / 2 + E(y.t)=g+(x. t)dB(t) (30) ! with price diffusion coefficient e(S.xT. .T. T)/ 2)g) + ( g) 2 "T "xT 2 "xT (24) has exactly the same Green function g(xT. The condition for a global Martingale is equivalent.xo. This shows that the risk neutral discounted price S=pe-rt with Sc=pc is drift free. (29) From this we obtain the corresponding price sde (with y=lnS(t)/Sc) ! dS = S 2 e(S. is a local Martingale.
2. Consider the price of a call for the case where p > pc: C(p. T. not just for p=pc. not returns.t)≠constant.t)≈x-m for x>>1. where f is the empirically observed distribution.4 Fat tails yield infinite option prices When R=constant then the reference price locates the maximum of f and the minimum of D [10. T " t) = er (t "T) $ (p T " K)g(xT . p is the known stock price at present time t<T. T " t) # er (t "T) % pe x x"µ dx = $ ln K/p c $ . x. T " t) = er (t "T) $ (p T " K)f(xT . C(p c . This is equivalent to evaluating C for the consensus price. K. so we get C(p c . or ‘value’ p=pc. One paper  starts correctly with fat tails in returns. T)dxT ln K/p c # . Acknowledgement . but then obtains a spurious Gaussian convergence factor by treating the stochastic variable u=x/tH in a time integral as if it would be both deterministic and drift free . so that one can then predict essentially any option price. (31) ! Here.0)=f(x.23]. Most make a simple mistake: they use fat tails in price  or price increments. But fat tails in price is an exponential distribution in x and this describes small to moderate returns. The evidence that traders don’t take fat tails into account is that the exponential returns model prices options in agreement with traders’ prices [7.12]. t)dxT ln K/p # . The same definition applies for R(x. (33) ! The option price will diverge for all p. (32) ! This is adequate for making our main point: the empirically observed density has fat tails in logarithmic returns [11. Many papers have been written purporting to price options using fat tails. We know the Green function analytically only for the case where g(x. K. not fat tails .0.9]. If one inserts a finite cutoff in (32) then the option price is very sensitive to the cutoff.t. K. f(x.t).
E.H. Dover: N. Physica A (2006). A. Cambridge. K.H. Cambridge. M. 383. Bassler. K. McCauley. and by TcSUH. 8. M. 2. Proc. Wax. tr. Brownian Motion and Martingales in Analysis. 2004. Wadsworth. Selected Papers on Noise and Stochastic Processes. McCauley & G. R. and Nonlinear Diffusion Equations. 422. 6. 3. Gunaratne & J. G. R. Gunaratne. R. 9.Y. Stratonovich.. A. 5. 10. McCauley. Mandelbrot & J. Dynamics of Markets: Econophysics and Finance. McCauley. 1992. Fla. 1963. 5848. J. Durrett.KEB is supported by the NSF through grants #DMR-0406323 and #DMR0427938. of SPIE conf. GHG is supported by the NSF through grant #PHY-0201001 and by TcSUH. Bassler.L. Stochastic Differential Equations. N. Silverman. van Ness. . I. 2006. and G.B. 1984. 2005. and thanks Vela Velupillai for helping us to spread our message within the economics community. Gordon & Breach: N. G. on Noise & Fluctuations 2005. Gunaratne. J. L. JMC thanks Cornelia Küffner for reading the manuscript and suggesting changes that made it and more readable. 178 (2003). L. Alejandro-Quinones.H.E. Physica A329.Y. 4.L. J. McCauley. Field. A. References 1.L.L. “Hurst Exponents.H. Török. Malabar. & J. Markov Processes.. Krieger. Timofeyev. Topics in the Theory of Random Noise. Nicol.L. Belmont.. SIAM Rev. 7. Gunaratne. L.131. by SI International and the AFRL.1968. W. Arnold. 2. 1954. 10. Physica 363A.
A. Picte. L. Markose and A. Elements of Partial Differential Equations. M. L. 229. R. Dacoroga. Political Economy 81. Belyushkin. in Dynamics of Complex Interconnected Systems.11. G. 2006. 2004. F. L. 1957.. E57. 21. 24.. 2002. T. 2006. 2001. B. McCauley.Y. Networks and Bioprocesses.Y. Harrison & D. Gunaratne. Physica A 331. K..2001. U. Chelsea. The Generalized Extreme Value (GEV) Distribution. Sneddon.V. Gunaratne. Springer. 1997.H. McCauley.E. ed. 13. J. Bassler. Ramazan. L. Rev. G. N. & K. 22. submitted. NY. Physica A237.H. An Intro.E. 16. . Müller. 23. Black and M. Bassler. Kaniadakis & P. Quarati.Y. Borland. by B.D. Gunaratne. Frank. The Theory of Probability. tr. 2006. & K. D. 637. Seckler. 415. 2005. and G. S. 405. and O. preprint. I.A. G. McCauley. 17. Phys. Alejandro-Quinones. Olsen. L. 18.L. 1967. N. Gnedenko. A. Alenthorn. J. Physica A 296.. Borland. Sneddon. 14. 19. Quantitative Finance 2. G. 1998. in preparation.T. 6634. Economic Theory 20. J. Scholes. 12. Kreps.M. V. 20. G. Kaniadakis. Implied tail Index and Option Pricing. M. 391. Bassler. Skjeltorp & A.E.B. 1973. J. Academic Pr. 1979. to High Frequency Finance.J.. 15. McGraw-Hill. N.H. 381.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.